Séminaires : Séminaire des Thésards

Equipe(s) : doctorants,
Responsables :Andrei Bengus-Lasnier, Eleonora Di Nezza, Ilias Ftouhi, Mario Gonçalves, Mahya Mehrabdollahi, Romain Petrides, Arnaud Vanhaecke
Email des responsables :
Salle :
Adresse :
Description

Le séminaire des thésards est l'occasion pour les doctorants de présenter des résultats et des problématiques dignes d'intérêt devant un public de non-spécialistes. L'ambiance y est informelle ; poser des questions naïves est encouragé, et les questions moins naïves sont bienvenues dans la mesure où elles n'entravent pas le bon déroulement de l'exposé.

Un mercredi sur deux à 17 h, en alternance entre Jussieu et Sophie Germain.


Orateur(s) Juan Pablo Vigneaux - ,
Titre Bethe entropy from a topological viewpoint
Date14/12/2017
Horaire18:00 à 19:00
Résume A graphical model is a graph that encodes interactions between random variables as a preferred factorisation of their joint probability distribution (the "global state"). They generalise several other models previously introduced in physics and machine learning, such as the Ising model and Bayesian networks. Several notions from statistical mechanics can be defined for these graphical models, most notably the entropy and the Gibbs free energy. These functions are difficult to compute, and Bethe introduced an approximation of the entropy that just depends on a "localised" version of the global state (its marginalisation over prescribed subsets of variables). This new function is easier to compute and gives in practice good results, even if people do not understand completely why. After introducing the concepts just described, we shall point to some topological developments (and open problems) around the Bethe approximation; for example, its interpretation in terms of graph coverings.
Salle
Adresse
© IMJ-PRG